Assessing the current landscape of AI and sustainability literature: identifying key trends, addressing gaps and challenges

Abstract

The United Nations’ 17 Sustainable Development Goals stress the importance of global and local efforts to address inequalities and implement sustainability. Addressing complex, interconnected sustainability challenges requires a systematic, interdisciplinary approach, where technology, AI, and data-driven methods offer potential solutions for optimizing resources, integrating different aspects of sustainability, and informed decision-making. Sustainability research surrounds various local, regional, and global challenges, emphasizing the need to identify emerging areas and gaps where AI and data-driven models play a crucial role. The study performs a comprehensive literature survey and scientometric and semantic analyses, categorizes data-driven methods for sustainability problems, and discusses the sustainable use of AI and big data. The outcomes of the analyses highlight the importance of collaborative and inclusive research that bridges regional differences, the interconnection of AI, technology, and sustainability topics, and the major research themes related to sustainability. It further emphasizes the significance of developing hybrid approaches combining AI, data-driven techniques, and expert knowledge for multi-level, multi-dimensional decision-making. Furthermore, the study recognizes the necessity of addressing ethical concerns and ensuring the sustainable use of AI and big data in sustainability research.

Introduction

Sustainability is becoming increasingly important, particularly following the establishment of the 17 Sustainable Development Goals (SDGs) by the United Nations (UN). The call for sustainability is now more resounding and pressing than ever. The 17 SDGs emphasize the necessity for global and local efforts to confront disparities, inequalities, and their far-reaching consequences on individuals and society [37, 42, 118]. They call for action to implement measures toward a fair and equitable world, focusing on addressing social, environmental, and economic dimensions. Technological advancements, such as the Internet of Things (IoT), Information and Communication Technology (ICT), blockchain, Big Data (BD), sensors, Artificial Intelligence (AI), and data-driven methods, are increasingly regarded as promising solutions for addressing sustainability-related problems. They can integrate various aspects of sustainability, enhance collaboration among stakeholders, improve resource management, expedite innovation, foster cross-industry applications, and enable long-term sustainable planning [62, 111, 154, 336]. A systematic and collaborative approach to integrating data, technology, and AI can mitigate social and economic inequalities and reduce ecological damage. This is achieved through the development of data-driven solutions, the creation of quantitative models to optimize energy and resource usage, the implementation of policies and regulations driven by informed decision-making, online monitoring of social, environmental, and economic indicators, and the optimization of economic activities to minimize carbon emissions [30, 336, 341].

A systematic approach founded on a thorough understanding of various sustainability-related topics is lacking to the best of our knowledge. Such an approach is important for creating a comprehensive and robust framework to integrate sustainable practices, policies, guidelines, regulations, and monitoring to help achieve sustainability objectives, such as embracing carbon–neutral technology and establishing a Circular Economy (CE). There are three primary research gaps: First, at the local level, there is a lack of understanding regarding the practical implementation of new technologies and the use of AI within institutions, local businesses, healthcare, and education. It encompasses addressing local challenges, resource management, and contributing to governance, all while reducing adverse impacts on communities. Second, at the global level, it is unknown how technology and AI can effectively integrate to address global-scale problems, such as establishing a resilient and sustainable Supply Chain (SC), adopting a CE, preserving biodiversity, and mitigating climate change [142, 147, 269, 294]. Third, integrating local and global sustainability concerns within an integrated strategy remains unclear. This strategy aims for global equity, inclusive growth across all sectors, and the contribution to a healthier and more collaborative world. In addition to these existing research gaps, another critical concern pertains to the integration of AI. These concerns encompass AI regulations, ethics, the digital divide, security, AI’s societal impact [97], and whether AI aligns with sustainability principles [73]. This alignment raises questions about whether AI contributes positively to sustainability or inadvertently accelerates resource depletion and reinforces biases.

To address these issues, this paper delves into the current state of AI and sustainability research, specifically focusing on the application of AI to address sustainability-related challenges at both local and global levels [276]. This exploration is carried out by answering the following research questions:

• 1) What is the state of research regarding AI and sustainability-related questions, trends, and collaboration between countries? A scientometric analysis of the literature sample addresses this question (see Ch. “Bibliometric Insights”).

• 2) What are the major implementation areas of technology and AI applications for sustainability? Topic modeling is performed to answer this question (see Ch. “Bibliometric Insights”).

• 3) How can different data-driven and AI-based methods utilized to address sustainability-related challenges be categorized? An analysis of the literature findings from Ch. “Bibliometric Insights” is conducted to tackle this question (see Ch. “AI, ML, and Data-Driven Methods”).

• 4) What are the dimensions of sustainable AI, considering its potential impact and implications? This question is answered by analyzing the literature results from Ch. “Semantic Analysis of Key Topics” through topic modeling (see Ch. “Sustainable Big Data and Analytics,” “Sustainable AI Characteristics and Challenges,” and “Sustainable Human AI Ecosystems”).

Addressing these questions improves understanding of AI’s status, potential, and challenges in promoting sustainable practices and solving multi-dimensional sustainability problems. This study offers a comprehensive overview of AI and sustainability literature (see Ch. “Publication Landscape Analysis”), the research collaboration among countries (see Ch. “Country-Wise Analysis”), and the current and emerging AI and data-driven approaches in sustainability research. It systematically explores AI methodologies applied across various sustainability applications (see Ch. “AI, ML, and Data-Driven Methods”), emphasizing the need for method integration to make decisions effectively and highlighting challenges with AI and its sustainable use (see Ch. “Sustainable Big Data and Analytics,” “Sustainable AI Characteristics and Challenges,” and “Sustainable Human AI Ecosystems”).

Background: artificial intelligence for sustainability

Over the past decade, advancements in AI have made significant strides toward effectively contributing to all facets of sustainability and addressing complex sustainability-related challenges. AI encompasses a broad spectrum of capabilities, with machines programmed to think and learn human-like cognitive abilities. These capabilities include environmental perception, information processing, decision-making, and taking action to achieve specific goals [277]. According to the European Union’s (EU) definition, AI entails “intelligent behavior that involves the analysis of the environment and the execution of actions, often with some degree of autonomy, to attain specific objectives” [334].

Realizing the potential of AI in achieving sustainable growth and meeting sustainability goals presents a challenge due to the multidimensionality of social, environmental, and economic sustainability issues. These challenges necessitate a comprehensive understanding of the interconnected nature of various problems [177, 280] and the collaboration of communities, nations, and diverse stakeholders. Governments, political and business leaders, innovators, scientists, and representatives from various industries worldwide are called upon to develop systematic efforts with a holistic and interdisciplinary approach to effectively address these challenges [300, 303, 342]. A collective endeavor holds the potential to foster sustainable practices in social and economic decision-making, energy consumption, resource utilization, manufacturing processes, and other sustainability-related objectives, all of which have long-term implications for overall societal well-being.

Furthermore, harnessing the potential of AI for sustainable growth and achieving sustainability goals requires several prerequisites: (1) Establishing a sustainable digital infrastructure; (2) implementing and developing robust and ethical AI solutions for both local and global contexts; (3) addressing the regulatory and sustainability concerns associated with AI; and (4) ensuring the development of environmentally friendly AI and preventing monopolization of AI resources and technology. It is essential to facilitate responsible AI practices, promote transparency, and consider AI implementation’s long-term social and environmental implications to maximize its positive impact.

Implementing AI methods poses various challenges, including their applicability, efficiency, performance, explainability, contextual understanding, associated risks, systematic frameworks for their application in sustainability tasks, the adaptability of their models in response to changing external factors, and user acceptability. Furthermore, the role of AI sparks essential discussions concerning its utilization in decision-making processes and the appropriate way AI should be employed. This leads to deliberations on whether AI should function as an assistant in decision-making or act autonomously as part of a multi-agent system. These concerns extend to AI-based decision-making within integrated systems, where multiple decisions across different integrated system components are exclusively determined by AI models, potentially leading to error and bias propagation. In such instances, the interrelated decision-making system could generate ripple effects that impact overall outcomes. Errors or biases originating at one stage may propagate throughout the system, influencing subsequent decisions and potentially resulting in incorrect or suboptimal results. Another essential aspect of utilizing AI is to address the trade-offs [112] associated with the three dimensions of sustainability (social, environmental, economic) at both global and local levels. This involves striving for an optimal balance between these trade-offs by leveraging AI to enhance decision-making. AI methods play a crucial role in the analysis of complex datasets, the prediction of outcomes, and the recommendation of solutions that maximize the benefits across all dimensions of sustainability [187].

Several studies focus on sustainability and AI-related research, particularly on efficient applications and challenges, including the fulfillment of SDGs [30, 70, 92, 341]. Kar et al. [152] present a systematic literature review on AI’s impact on the sustainability of technical challenges, social issues, and environmental causes. Falk and van Wynsberghe [90] propose three criteria for ensuring the appropriate application of AI for sustainability: “Monitoring and information provision, sustainability analysis of the application, and an action component contributing to a sustainability goal.” Nishant et al. [232] discuss the challenges and limitations of AI for sustainability, emphasizing environmental governance, industrial environmental performance, and risk reduction. Galaz et al. [97] examine the implications of AI interacting with society as a socio-technical system and using responsible AI. Khakurel et al. [159] conducted a study on the long-term impact of AI on sustainability in various dimensions in a previous literature search along with a focus group study. S. R. Wu et al. [353] discuss the impact of AI and other technologies in smart cities, SCs, and energy systems. Kopka and Grashof [172] conducted an empirical study to investigate the impact of AI on energy consumption and found that AI has the potential to both conserve and consume energy.

Various concerns related to the effective use of AI must be considered and addressed. These include AI integration, regulations, governance, the digital divide, potential threats, security, and the overall value derived from AI [97, 155]. It is imperative to assess whether AI effectively contributes to sustainability goals or if AI-based decisions unintentionally undermine these objectives. For instance, a critical evaluation of AI-based decisions should consider whether they result in the depletion of resources, favor powerful entities due to inherent design biases, or introduce contradictions in sustainability tasks. It is paramount to ensure that AI does not facilitate monopolies through biased decisions or intentional design aimed at monopolistic advantages. These pivotal issues must be addressed to guarantee that AI outcomes align with sustainable principles and benefit society [73]. Several literature reviews are available on the topic of AI and sustainability, but they are mostly focused on specific aspects such as sustainability in manufacturing [131, 134], urbanization and city planning [175], and smart and sustainable farming [11]. However, no research on the overall landscape of AI and sustainability research is currently available. A topic modeling and text analysis approach is used to study a broader range of sustainability research to address this gap.

Methods

The workflow to comprehend the current landscape of AI and sustainability research is divided into three steps and visualized in Fig. 1. First, a literature search was conducted using specific search strings that match the titles or keywords of the literature. In the second step, data and text analysis, we conducted four types of analysis: (1) A publication landscape and trend analysis was completed. (2) A country-wise analysis of countries emphasizing sustainability and AI research, as well as collaborations among countries, was performed. (3) A keyword analysis of the searched literature examined different sustainability aspects in co-occurrence network modules. The focus was identifying significant keywords and a network-based approach for analyzing the interconnection between sustainability, AI, and technology terms (co-occurrence networks). Module identification in co-occurrence networks of keywords and enrichment analysis was involved. (4) A semantic analysis of key topics involved topic modeling using the BERTopic model to distinguish key topics within the searched literature and clustering identified topics and representative papers. The results obtained in steps (3) and (4) were utilized to explore various AI and data-driven methods and models applied in diverse social, environmental, and economic sustainability applications. Additionally, they were used to examine the topics that underscore fundamental AI-related concerns relevant to sustainability. Third, we summarized the results and discussed key methods and sustainable AI challenges.

Fig. 1

A Schematic View of the Literature Analysis

Literature search

A comprehensive search in the Scopus database, which offers access to high-quality publications spanning a wide spectrum of scientific, engineering, humanities, and business disciplines, was conducted. This database provided ample literature from key sustainability journals for the scientometric analysis. Although searching in additional databases might have augmented the search, we focused exclusively on Scopus due to the limitations of access and processing speed associated with other databases our university does not subscribe to. The following search string was used to search the title or keywords of the documents:

s = (sustainability OR sustainable) AND (“artificial intelligence” OR ai OR “machine learning” OR “data driven” OR data OR analytic OR forecast OR algorithm OR optimization OR “data mining”)

The search was not restricted by time period, publication type, or scientific discipline. The only requirement was that the documents must be in English. In total, a sample of 1,982 documents was collected, each containing various types of metadata, including title, abstract, affiliations, and keywords. The search was restricted to the document titles and keywords aiming to collect documents that primarily focus on issues related to sustainability, particularly those that involve the applications of AI and Machine Learning (ML). The search was intended to ensure the selected documents are closely aligned with the intersection of sustainability problems and AI and ML technologies independent of any specific research area utilized. From the obtained document keywords, those conceptually linked with the three aspects of sustainability were extracted. The authors selected these keywords manually, which are important for the broader understanding of the application of AI and data-driven methods for sustainability. The selected keywords are shown in Table 1.

Table 1 Key terms extracted from the keywords that reflect the social, environmental, and economic aspects of sustainability

Keywords significance analysis

The keyword significance analysis is conducted to identify the most significant keywords with higher frequencies, ensuring their appearance is not random among the selected documents. Let a document set $D$ with $\left|D\right|$ documents and keyword set $kw=k{w}_{{d}_{1}}\cup k{w}_{{d}_{2}}\dots \cup k{w}_{\left|D\right|}$.

Let the $M={\left[{m}_{i,j}\right]}_{i=\mathrm{1,2},\dots ,\left|kw\right|}^{j=\mathrm{1,2},\dots ,\left|D\right|}$ be a binary matrix (${m}_{ij}\in \{\mathrm{0,1}\}$) representing the presence or absence of keywords, from $kw$ set in document set $D$. The frequency of a keyword $k{w}_{i} \in kw$ is ${f}_{k{w}_{i}}={\sum }_{j=1}^{\left|D\right|}k{w}_{i}$. The following null hypothesis was tested:

$$H0:f_{{(kw_{i} )}} \, \le \,f_{{(kw_{i} )}}^{random}$$

$$H1:f_{{(kw_{i} )}} > \,f_{{(kw_{i} )}}^{random}$$

To test the null hypothesis, matrix entries were randomized, $M$, and obtaining ${M}_{randomized}^{b}$ where $b=\{\mathrm{1,2}, \dots B=\mathrm{5,000}\}$. For ${M}_{randomized}^{b}$, ${f}_{k{w}_{i}}^{randomized}$ was calculated for $k{w}_{i}$. The p-value to test the significance of ${f}_{k{w}_{i}}$ was calculated as follows:

$$p\left(k{w}_{i}\right)=\frac{\# {\{f}_{k{w}_{i}} > {{f}_{k{w}_{i}}^{randomized}\}}_{b=1}^{B}}{B}$$

Further, $\left|kw\right|$ hypotheses were tested. Therefore, multiple testing corrections False Discovery Rate (FDR) were applied to control the false positive results in keyword significance analysis.

Network analysis

In this analysis, network-based approaches were applied for analyzing keyword-co-occurrence analysis and country collaboration analysis by constructing networks, identifying network modules, and performing module enrichment analysis for the keyword co-occurrence network.

Network construction

For constructing the network, the Jaccard index was used to measure the similarity between two keywords, quantifying how often these two keywords appear together in a set of documents. The Jaccard index was utilized to establish both keyword co-occurrence and collaboration networks. The method follows network modules and performs module enrichment analysis for the keyword co-occurrence network. Let a document set $D$ with $\left|D\right|$ documents and keyword set $kw=k{w}_{{d}_{1}}\cup k{w}_{{d}_{2}}\dots \cup k{w}_{\left|D\right|}$. A keyword-document matrix $[{{\text{m}}}_{{\text{ij}}}{]}_{i=\mathrm{1,2},\dots \left|kw\right|}^{j=\mathrm{1,2},\dots ,\left|D\right|}$ was created, and ${m}_{ij}\in \{\mathrm{0,1}\}$. The value of ${m}_{ij}=1$ if the keyword ${k}_{i}$ is mentioned in the document ${d}_{j}$ in their keywords. The Jaccard index measures the similarity between the two keywords.

$$ji\left({kw}_{a},{kw}_{b}\right)=\frac{||{m}_{{kw}_{a},.}\cap {m}_{{kw}_{b},.}{||}_{/\left||\mathrm{0,0}\right||}}{||{m}_{{kw}_{a},.}\cup {m}_{{kw}_{b},.}{||}_{/\left||\mathrm{0,0}\right||}}$$

Similarly, a country-based analysis was created for computing similarity based on their common co-occurrence by constructing matrix $[{m}_{ij}{]}_{i=\mathrm{1,2},\dots \left|c\right|}^{j=\mathrm{1,2},\dots ,\left|D\right|},$ and ${m}_{ij}\in \{\mathrm{0,1}\}$, where $D$ is a document set with $\left|D\right|$ documents and $c={\{c}_{1},{c}_{2},\dots ,{c}_{\left|c\right|}\}$ are $\left|c\right|$ countries.

$$ji\left({c}_{a},{c}_{b}\right)=\frac{||{m}_{{c}_{a},.}\cap {m}_{{c}_{b},.}{||}_{/\left||\mathrm{0,0}\right||}}{\left||{m}_{{c}_{a},.}\cup {m}_{{c}_{b},.}\right|{|}_{/\left||\mathrm{0,0}\right||}}$$

The keyword co-occurrence and country collaboration network are created as follows: Let $G$ be an undirected weighted graph $G=\left(V,E,w\right), w: E \to R$ where the $V$ are the vertices and $E$ are the edges. The number of vertices is equal to the keywords $\left|kw\right|$ (or country $\left|c\right|$), and

$$E\left( {i,j} \right) = 1,{\text{if}}\,ji\left( {{\text{kw}}_{a} ,{\text{kw}}_{b} } \right)\, > \,\alpha \,\left( {ji\left( {c_{a} ,c_{b} } \right)\, > \,\alpha \,for\,country\,collaboration} \right)$$

$w\left(i,j\right)=ji\left({kw}_{a},{kw}_{b}\right)$($ji\left({c}_{a},{c}_{b}\right) for\,country\,collaboration\,network$).

Module detection

We utilized the multi-level modularity optimization algorithm developed by Blondel et al. [44] to partition the constructed weighted keyword co-occurrence and country collaboration networks. This algorithm groups nodes that exhibit greater proximity and similarity compared to other nodes. The multi-level algorithm outperforms alternative module detection algorithms, offering faster processing times and superior results.

An adjusted modularity optimization approach was employed for the country collaboration network using multi-level module detection to optimize the modules. In this approach, a small fraction $\left(v=0.0001\right)$ from edge weights, i.e., $w\left( {i,j} \right) = max\left( {w\left( {i,j} \right){-}v,0} \right)$ was iteratively subtracted. The multi-level module detection algorithm for weighted graphs was run until the modularity reached its maximum value or until 2% of the total edge weights became zero. This approach prioritizes countries into the same modules, which shows stronger collaboration and emphasizes stronger collaboration based on the Jaccard index value.

Module enrichment analysis

Module enrichment analysis is used to identify modules in the keyword co-occurrence network that are enriched for a specific set of keywords or annotations (in this case, different aspects of sustainability). This can be useful for understanding the characteristics of different modules in the co-occurrence network and identifying important modules for a particular phenomenon. Module enrichment analysis was applied to identify whether various sustainability keywords were significantly represented in a module obtained from the keyword co-occurrence network $(G)$. The following hypotheses were tested:

H0: The presence of observed keyword pattern expressing sustainability aspect ${s}_{a}$ in module ${m}_{i}$ is by chance.

H1: The observed number of keywords expressing sustainability aspect ${s}_{a}$ in module ${m}_{i}$ is significantly higher than the expected number by chance.

The Fisher exact test was applied to test the hypotheses, comparing the significance of the association between two sets. In this case, we evaluated the association between sustainability aspects and modules identified in the keyword co-occurrence network by assessing the presence or absence of keywords expressing a particular sustainability aspect in a module of the keyword co-occurrence network. The contingency table illustrating the presence and absence of the keyword pattern (as shown in Table 1) of each sustainability aspect in a module is presented in Table 2, described as follows:

Table 2 Contingency table

Let ${\text{N}}=\left|V\left(G\right)\cup s\right|$ be the total keywords representing the union of vertices of the co-occurrence network and all terms from the sustainability aspects (see Table 1), where $\{{{\text{s}}}_{{\text{social}}}, {{\text{s}}}_{{\text{environmental}}}, {{\text{s}}}_{{\text{economic}}}\}$ are sustainability aspects and $s={s}_{social}\cup {s}_{environmental}\cup {s}_{economic}$, and a module ${m}_{i}\subset V\left(G\right)$. The values of the contingency table are defined as:

$${{\text{N}}}_{11}:\text{are the number of pattern matches of keywords from s}_{{\text{a}}}\text{in m}_{{\text{i}}}$$

$${{\text{N}}}_{12}:\text{are the number of keywords in s}_{{\text{a}}}\text{ do not match with keywords in m}_{{\text{i}}}$$

$${{\text{N}}}_{21}:\text{are the number of keywords in m}_{{\text{i}}}\text{ do not match with keywords in s}_{{\text{a}}}$$

$${{\text{N}}}_{22}:\text{are the remaining keywords not present in m}_{{\text{i}}}\text{ and not with keywords in s}_{{\text{a}}}$$

The test calculates the probability of obtaining the observed outcome, assuming the null hypothesis is true. The probability of ${N}_{11}$ sustainability aspect keyword pattern matches with keywords in a module:

$$p=\frac{\left(\genfrac{}{}{0pt}{}{{N}_{1.}}{{N}_{11}}\right)\left(\genfrac{}{}{0pt}{}{{N}_{2.}}{{N}_{21}}\right)}{\left(\genfrac{}{}{0pt}{}{N}{{N}_{.1}}\right)}$$

The probability $p$ follows a hypergeometric distribution, representing the likelihood of observing an overlap of sustainability aspect-related keyword patterns within the module when randomly drawing keywords from the data.

Semantic analysis

Topic modeling is employed for the semantic analysis of publication texts, combining the title, keywords, and abstract. For the topic modeling of the obtained literature, the BERTopic [107] approach was applied to identify key research themes related to sustainability. For the topic modeling, our objective is to explore topics that can be comprehended best through semantic analysis while maintaining diversity and coherence. We have opted for BERTopic as our method of choice to achieve this goal due to its effectiveness with large corpora. Several studies have demonstrated that BERTopic outperforms other methods in capturing contextual information and generating coherent topics. Furthermore, BERTopic has proven effective in various assessments compared to different topic modeling techniques [84, 98, 332].

The BERTopic approach follows four main steps: (1) Generating embeddings, (2) dimensionality reduction, (3) clustering, and (4) c-TF-IDF-based topic extraction from the clusters. The “sentence-transformers” library transforms the input text into numerical representations (embeddings) optimized for semantic similarity. The “all-minilm-l6-v2” sentence transformer model, which maps the input text of each document into a 384-dimensional numerical vector, was used. The model then applies Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction $(n=20)$ and uses hdbscan for clustering, selecting the number of clusters. The “BERTopic” python package automatically generates the topic representation through c-TF-IDF-based extraction of the most important keywords from the clustering solution. Additionally, the coherence was enhanced, and stopwords were reduced in the extracted topics by using the “KeyBERTInspired” representation model in BERTopic. From the topics identified by BERTopic, hierarchical Ward clustering was applied to the topic-term c-TF-IDF matrix to find clusters of common topics resulting from the topic modeling process.

Topic optimization

We utilize the topic coherence measure Coherence (C_V) to optimize and fine-tune topics. The measure involves varying the parameters of the number of neighbors [3, 5, 8, 10, 15, 20, 25, 30] for the “UMAP” function and the minimum cluster size parameter for “hdbscan” clustering (from 3 to 30). We perform topic modeling for each combination, where the BERTopic hdbscan automatically selects the number of clusters and computes the C_V coherence [268]. The results are evaluated based on the C_V score and the number of topics obtained from the topic modeling. Additionally, we manually evaluate the topic keywords to assess if the topics identified by the selected model uniquely represent themes in sustainability and AI research.

Significance analysis of projection of the clustered embeddings

For comparing differences between clusters obtained by the hdbscan algorithm in “BERTopic.” We analyze the UMAP-reduced 2D projection of document embeddings to assess the statistical significance of clusters. We applied the non-parametric $N$-statistic method [35] to test the null hypothesis regarding the equality of distributions of data points in two clusters. The method is described as follows: Let $X$ and $Y$ be the p-dimensional UMAP reduced embeddings of documents in clusters $i$ and $j$ with samples $m$ and $n$, respectively, and ${F}_{C{L}_{i}}$ and ${G}_{C{L}_{j}}$ are these clusters’ distribution functions. The null and alternative hypotheses are expressed as follows: $H0:{F}_{C{L}_{i}}={G}_{C{L}_{j}}$,${\text{H}}1:{F}_{C{L}_{i}}\ne {G}_{C{L}_{j}}$

The $N$-statistic for the comparison is defined as follows:

$$T_{{m,n}} = \frac{{mn}}{{m + n}}m\left[ {\frac{1}{{mn}}\sum\limits_{{i = 1}}^{m} {\sum\limits_{{j = 1}}^{n} {\left| {\left| {X_{i} - Y_{j} } \right|} \right| - \frac{1}{{2m^{2} }}} } \sum\limits_{{i = 1}}^{m} {\sum\limits_{{j = 1}}^{m} {\left| {\left| {X_{i} - X_{j} } \right|} \right| - \frac{1}{{2n^{2} }}} } \sum\limits_{{i = 1}}^{n} {\sum\limits_{{j = 1}}^{n} {\left| {\left| {Y_{i} - Y_{j} } \right|} \right|} } } \right]$$

Bibliometric insights

Publication landscape analysis

The literature search obtained 1982 papers, including research articles, review papers, books and book chapters, conference papers, and editorials. Table 3 shows the different journals that publish papers on sustainability and AI-based applications.

Table 3 Number of papers of top 30 journals

The largest fraction belongs to other categories, including a mix of conference papers, journal papers, and others, but no category exceeds the top 30 journals. Tables 4 and 5 display the categories of top review papers, research articles, and collaborations based on their citations. The top articles belong to various subject areas or topics of sustainability and are sourced from different journals.

Table 4 Top papers based on citations (A and B) and number of collaborations (C)

Table 5 Papers ordered based on collaborations

Figure 2 shows the number of publications per year. The earliest paper is from 1995 and discusses the “community option model” to provide the outcome of policies and actions for community development and management. The subsequent years show a slow growth in the number of publications until 2017; afterward, the rise in the number of publications is higher. Just after the year 2018, the frequency of publication is in triple digits, showing the maturation and application of data-driven technology. AI has reached a level where it can be utilized systematically with expertise and necessary resources for various applications and multiple disciplines with sustainability-related topics at the center.

Fig. 2

Number of publications over the years related to sustainability and AI

The next result is the representation of social, environmental, and economic aspects in the current literature’s keywords. For this analysis, all keywords from the literature were extracted, duplicates eliminated, and keywords (including open compound words) and subparts of compound words relevant to the general concept of sustainability were manually selected. These selected keywords are categorized into one of three aspects, as presented in Table 1. Thus, the social, environmental, and economic aspects were defined. These three categories of words are used to search for patterns in the titles and keywords.

The proportion of different aspects of sustainability using keywords in Table 1 is as follows: Let us consider a set of documents $D={\{d}_{1},{d}_{2},\dots$} with $\left|D\right|$ documents. There is also a set ${k}_{a}=\{k{w}_{1},k{w}_{2},\dots \}$ with $\left|{k}_{a}\right|$ words related to aspect $a$. Let $P\left(k{w}_{i},{d}_{j}\right)$ be a pattern search function that returns 1 if it finds a match for $k{w}_{i}$ in ${d}_{j} (keywords)$, and 0 otherwise. The proportion for aspect $a$ is calculated as follows:

$${\text{Total count for aspect a in D: a}} = \sum\limits_{{i = 1}}^{{\left| {k_{a} } \right|}} {\sum\limits_{{j = 1}}^{{\left| D \right|}} {P\left( {kw_{i} ,d_{j} } \right)} }$$

It was searched for patterns related to each aspect in the titles and keywords of documents for different years. The year-wise counts for each aspect of sustainability are presented in Fig. 3. There has been a consistent upward trend in publications related to sustainability, AI, and data-driven approaches. The keywords in these publications have been categorized into social, environmental, and economic groups, and there has been a proportional increase in the count of keywords in all three categories. The results reveal a consistently higher count of environmental-related keywords since 2004.

Fig. 3

Frequency of the terms related to the triple bottom line of sustainability (social, economic, and environmental)

In contrast, social-related and economic-related keywords have been less prominent than environmental-related ones, indicating a lower representation in the papers. Until 2015, they were closely matched in the count. However, from 2016 onward, there has been a significant and consistent increase in the count of economic-related keywords compared to social-related keywords. The social dimension is complex, and implementing AI in this area presents a substantial challenge, necessitating greater expertise, domain knowledge, investments, systematic data collection, and implementation trials that benefit society. The increasing number of publications underscores the subject’s importance, emerging trends, and diversity within the field.

Country-wise analysis

The initial step in analyzing global research on AI-based applications for sustainability involves examining the country affiliations of the authors who have contributed to the research papers. This examination provides insight into the scope and prevalence of sustainability research. In cases where multiple authors are listed on a paper and share the same country affiliations, they were treated as a single instance for that specific country. This approach streamlines the analysis by ensuring each paper is associated with a unique country affiliation, thereby preventing potential bias. Table 6 displays the leading countries with a minimum of 20 publications among the 102 countries with at least one publication.

Table 6 Country-wise publication distribution

In these results, institutions or universities in the United States are affiliated with the most publications by authors, followed by China, India, the United Kingdom, Germany, and other countries. One interesting observation is that these are larger economies and higher-income countries investing in sustainability-related research. Next, the number of publications per million population was compared to the Gross Domestic Product (GDP) based on Purchasing Power Parity (PPP). Population and GDP (PPP) data for 2022 were gathered from the World Bank website (databank.worldbank.org). This data could only be obtained for 91 countries. The visualization of GDP (PPP) on the x-axis and the logarithmic value of publications per million population on the y-axis is presented in Fig. 4. This visualization highlights the focus on sustainability-related applications and research. Countries such as Norway, Ireland, Finland, Denmark, Sweden, Cyprus, and Macao, which fall within the high to middle range of GDP (PPP), exhibit higher publication rates per million population. In contrast, lower-income or densely populated countries tend to have significantly lower publication rates per million. This consideration is particularly important for implementing locally relevant sustainability initiatives.

Fig. 4

GDP per capita vs number of publications per million

The finding pertains to country-wise publication trends, revealing that countries with more advanced industrial economies or higher per capita incomes tend to exhibit more publications. This could be attributed to the rise of Industry 4.0 (I4.0), research investment, adopting and addressing sustainability, especially for economic and environmental dimensions, alternative energy sources with less carbon footprint, and the UN’s SDGs. However, comparing publications per million population versus GDP per capita reveals that many larger economies lag significantly behind, which is also the case for underdeveloped countries. This analysis does not reflect the actual status but only reflects AI and sustainability research papers from the Scopus database. AI and sustainability research must be emphasized to understand various local sustainability challenges, particularly aligning with the demography. Otherwise, the full potential of AI-based approaches for sustainability may remain underutilized. Demography and population growth are important factors when investing in sustainability-related research and applications. Bhargava [39] discusses the implications of rapid population growth and emphasizes its inclusion in sustainability research.

Next, the collaboration network from the country-document binary matrix using the Jaccard index for collaborative analysis was constructed. As discussed in the method section, countries with at least three collaborations and a Jaccard index value greater than $\alpha >0.01$ were selected. Next, module detection algorithms were applied. Seven cooperation modules were obtained from the module detection analysis. The module of countries’ collaborations is shown in Fig. 5. Countries collaborate with their specific module and with other countries. The relative weight measured as the Jaccard index is higher for countries within the same module.

Fig. 5

Collaboration among 70 countries related to sustainability and AI research

Sustainability is interdisciplinary and necessitates collaboration on various issues, particularly regional and global ones. The modules identified from collaboration networks of countries based on the Jaccard index reveal three main findings: the collaboration among large economies, the collaboration among European countries, and the collaboration among neighboring or closely situated countries. The collaboration network consists of only 70 countries, as a minimum of three collaborative publications had to occur for inclusion. The major emphasis is understanding how the AI applications of sustainability-related research are distributed worldwide. As several studies have highlighted, overcoming the digital divide and lack of infrastructure and human resource expertise in underdeveloped countries requires investment in research, education, and collaboration to enhance research in these regions. Achieving social, environmental, and economic sustainability is equally important and a priority for underdeveloped nations. With sufficient research on the social, environmental, and economic dimensions and a comprehensive understanding of the system, the advantages of sustainability tasks will be achievable only through collaborative efforts [96].

Keyword analysis

The keyword analysis highlights the significance of various technologies in conjunction with sustainability concepts, their application areas, and their interrelationships. The keyword analysis began with a pool of 5723 unique keywords. When the significance analysis was applied to identify the most meaningful keywords, it narrowed them to 67 (see Table 7). These keywords encompass sustainability-related concepts and technology/data-driven methods.

Table 7 Top significant keywords

Table 13 (see Appendix) displays the keywords in each of the 14 modules detected in the co-occurrence network. Enrichment analysis was conducted to identify sustainability-related aspects within these modules. This analysis revealed that 11 out of the 14 modules were enriched with at least one aspect related to sustainability. Figure 6 illustrates the co-occurrence network and labeled nodes with a degree greater than 20.

Fig. 6

Keyword Co-occurrence and Strong Connections

Table 8 shows the outcomes of the enrichment analysis, highlighting the significant modules that contained enrichment for at least one aspect. Notably, the findings indicate that not all modules were uniformly enriched with all three aspects (social, environmental, and economic). Seven modules significantly represent the environmental and economic aspects. The key terms in Table 8 (column 4) represent key sustainability terms. The results reveal variations in the distribution of sustainability-related aspects throughout the network of keywords.

Table 8 Identification of significant representation of sustainability-related terms in various modules using the Fisher test

The keywords and co-occurrence analysis focus on two important aspects: First, identifying common key terms, which highlight a central subject’s association with other key terms, and second, identifying current and emerging trends. The results shown in Table 7 highlight keywords that emphasize sustainability-related themes and keywords that emphasize technology and data-driven methods. The top two keywords can be understood in terms of sustainability or sustainable topics with AI applications. “Sustainable development” is a major abstract keyword that emphasizes that significant publications related to the search criteria actively consider or try to align with the principles of sustainable development and demonstrate the widespread recognition of the importance of sustainability in that field of research. However, the global sustainability objectives can conflict with the various objectives of these studies.

Further research is warranted where each work addressing sustainability is critically evaluated for global sustainability questions. Keywords such as smart and sustainable cities, I4.0, renewable and sustainable energy, and agriculture have gained considerable attention by being discussed by different research groups. However, this trend could lead to emerging issues if trade-offs exist with global sustainability goals. The keywords covering data-driven technologies show that “decision support system” is the key term followed by BD, IoT, Deep Learning (DL), neural network, sustainable AI, optimization, blockchain, and cloud computing. Other keywords show the importance of using technologies and data-driven methods for sustainability for effective and complex decision-making, real-time data-driven solutions, and higher prediction accuracy. They address trade-offs, transparent and traceable communication and transactions, and scalable, on-demand computing resources and services. These technologies and methods can be useful instruments for systems integration [197].

Analyzing the keywords’ co-occurrence network indicates the relationships between various keywords. The module detection algorithm applied for the co-occurrence network shows a high modularity score of 0.58, indicating that the keywords form distinct groups closely connected within themselves and loosely connected to other groups. These groups represent different research topics and are composed of technology-related, sustainability-related, and other keywords. These keywords highlight the relation between various topics, research themes, and subjects discussed in each group’s topic. For instance, Module 1 (see Table 13, Appendix) includes, amongst others, the keywords BD, IoT, I4.0, SDGs, CE, business model, and strategies. Economic sustainability, business model development, manufacturing, and the SDGs are not independent but intertwined. The modules highlight the major interconnected themes of research topics related to sustainability, AI, and data-driven methods.

The enrichment analysis using the Fisher exact test determines which of the 14 modules were significantly enriched toward social, environmental, or economic dimensions of sustainability. Eight modules were significantly enriched to at least one aspect of sustainability. Modules 1, 5, and 11 showed topics interconnected with economic and environmental aspects. Module 2 showed the interconnection of social and economic dimensions. The remaining modules were enriched with economic (three modules) and environmental (three modules) dimensions. Only Module 2 was enriched with the social dimension, specifically regarding sustainable AI implementation, regulation, governance, and protecting privacy. When gaining economic benefits from AI and data-driven models, these are important social aspects. The enrichment of the environmental dimension may be due to the selection of more environmental keywords, which could be different if more keywords related to sustainability for different aspects had been obtained. However, the current analysis overviews topics enriched with specific aspects within each module. This analysis can be useful for understanding different research topics and the extent of involvement in the three dimensions of sustainability within the specific sustainable theme, identifying all three sustainability-related concerns, and addressing all aspects equally.

Semantic analysis of key topics

To determine the optimal number of topics, we evaluated 64 models using different combinations of UMAP’s number of neighbors parameter and hdbscan’s minimum cluster size parameter, calculating the corresponding C_V scores. The results of these evaluations are presented in Fig. 7. The figure shows that the C_V coherence increases as the number of topics decreases, while the score decreases for higher numbers of topics. The selected input parameters for BERTopic influence the variation in C_V scores. The top three C_V scores exceeding 0.47 correspond to models with topics ranging from 1 to 3. However, these models are deemed suboptimal, as they do not exhibit sufficient variation in topics related to sustainability and AI.

Fig. 7

Number of topics vs. C_V score obtained by varying parametes of different BERTopic models

In contrast, the fourth and fifth largest C_V scores, 0.476 and 0.448, are associated with models having 12 and 13 topics, respectively, displaying less variation in topic distribution. The sixth-largest C_V score, 0.446, is the selected model with 34 topics (highlighted in blue). We chose this model because its coherence is relatively close to the fourth and fifth models, capturing a larger variation in AI and sustainability topics. The selected model has 34 topics and is characterized by UMAP’s number of neighbors set to 8 and hdbscan’s minimum cluster size set to 10. The topic modeling with BERTopic generated 34 clusters of keywords, which are shown in Table 11 (see Appendix). The description of these clusters is shown in Table 9. Through topic modeling, key research themes emerging from multiple studies are identified. While many papers do not align with any specific theme, these papers do mention AI and sustainability.

Table 9 Different topic labels identified using BERTopic (details of topic keywords are shown in Table 11 in the Appendix)

Cluster -1 contains 643 documents that do not belong to any specific cluster. This occurrence can be attributed to two primary reasons. First, it might be due to a higher noise level in the data, causing these documents to not fit neatly into any theme or category. Second, the clusters formed may be smaller than the minimum cluster size parameter, set at ten during the implementation. Comprehensive parameter tuning may be required to address this issue and assign these unclustered documents to appropriate clusters. However, it is important to note that fine-tuning these parameters may provide more granular topics, but results remain unaffected. Figure 9 (see Appendix) presents a two-dimensional projection of document embeddings obtained through the UMAP approach. This visualization highlights the distribution of all the documents across different clusters resulting from the BERTopic modeling. Notably, the grey-colored points represent the documents in the -1 labeled group scattered throughout the visualization, indicating their lack of clear association with any specific cluster. Using a multivariate hypothesis test using N-statistic, we compared the clusters (topics) obtained from “hdbscan” in BERTopic using 2D UMAP projection (Figure 9, see Appendix) of document embeddings. The aim was to determine if significant differences existed between the topic distributions shown by the two main UMAP components. The results of the $N$-statistic are shown in Figure 10 (see Appendix). The distribution of different clusters is shown to be significantly different, showing that the identified clusters, as represented by topics, are distinct. The p-values between all pairs of clusters are < 0.00001. In the final step of the analysis, hierarchical clustering (ward) was performed on the topic-text c-TF-IDF matrix. First, cosine similarity between topics was calculated, cosine dissimilarity as a distance measure was obtained, and ward clustering was applied.

The $C$-index was used to optimize the number of clusters [124]. Six clusters from 33 topics were obtained (see Fig. 8). Cluster 1 (from left) contains only one member, Topic 32. The second cluster has three members: data-driven business model innovation (BMI), sustainable SC, and sustainable supply chain management (SCM). The third cluster is the biggest one, with 13 members, and the main topic is related to AI applications in different areas, highlighting the closer connection in terms of semantic similarity of the text. The remaining clusters have eight, four, and four members, respectively.

Fig. 8

Topic clustering based on topic-terms c-TF-IDF matrix with cosine dissimilarity (C-Index optimized clusters)

To gain insight into the details of the research themes, the topic modeling approach using BERTopic identified 33 different sustainability-related topics (see Table 9 and Table 11, Appendix). These categories include sustainable development, renewable energy, manufacturing, health care, urban planning, product sustainability, sustainable SC, education, and other topics. The identified topic outlines the research and efforts and major emerging research themes of sustainability.

One of the results is related to the sustainable development of nanomaterials and nanotechnology. Nanomaterials and nanotechnology have widespread applications and are promising technologies offering disruptive transitions in the use of resources, environmental sustainability, transportation, energy usage, health, and other areas [257]. AI and nanotechnology can offer promising results [103, 270], highlighting the increasing importance of sustainable practices and innovations in nanoscience and nanotechnology, emphasizing sustainability’s growing intersection and cutting-edge technological advancements.

Additionally, by conducting further clustering of the topic-term matrix (c-TF-IDF matrix) to find the clusters of topics, five distinct groups from the initial 33 topics could be distinguished. These clusters of topics share interconnected themes within a broader context. For instance, the second cluster from the left (colored green) comprises three topics: Topic 25 (focused on BD and BMI), Topic 8 (centered around sustainable SC), and Topic 17 (on sustainable SCM). These topics are closely related, particularly regarding their underlying theme of adopting data-driven approaches for sustainable BMI.

Discussion

The data analysis provides insights into current trends in AI and sustainability, covering three aspects: sustainability trends, the emphasis on AI and sustainability research by country, and keyword analysis highlighting key research themes using AI and data-driven methods. Additionally, topic modeling is utilized to identify key topics within sustainability research. The following sections will discuss the various methods applied to sustainability and AI. The analysis of topics, keywords, key application areas of AI, and the intersection of technology and AI in the context of sustainability emphasizes four key aspects that impact the role of AI in achieving sustainable outcomes and shaping the future of sustainable AI practices. The first aspect is various data-driven and AI methodologies applied for different levels of sustainability-related problems. It is important to understand the scope and application of these methods, as well as the integration of these methods, which is essential for informed and effective decision-making for multi-dimensional and multi-level sustainability problems. The second aspect revolves around the scope and challenges associated with data-related characteristics, BD, and BD technologies, particularly its sustainable utilization. The third aspect centers on adopting sustainable AI practices, highlighting the importance of integrating sustainability principles into AI development and deployment. Lastly, the fourth aspect emphasizes facilitating sustainable human-AI ecosystems, where the widespread use of AI and the active interactions of humans and AI systems lead to the emergence of different unknown social, environmental, and economic phenomena. These four aspects collectively define the trajectory of AI’s contributions to sustainability and illustrate the path towards the future of sustainable AI practices. These aspects will be briefly discussed in the next sub-chapters.

AI, ML, and data-driven methods

The AI and data-driven methods for sustainability applications are divided into nine categories: DL and neural network models; Other supervised and unsupervised ML models; forecasting methods; optimization methods; fuzzy methods; multi-criteria decision-making; complex network-based analysis, ontology-based models; statistical, stochastic, Bayesian methods. These method classes are based on the surveyed literature and keyword analysis. The different categories of the methods and relevant publications are shown in Table 10 (additional details in Table 12, Appendix), and the corresponding plausible sustainability dimensions they address.

Table 10 Different ML and data-driven methods applied to different sustainability problems

The recent publications published within the last eight years were selected. The different categories of sustainability dimensions are based on their relevance, as mentioned in the abstract and keywords. Some publications can belong to more than one category regarding methods and sustainability criteria. However, it was assigned to only one category based on the importance of the subject given in the article. These models are applied across a wide range of sustainability domains. However, the studies might not directly relate to the 17 SDGs but rather focus on the general idea of sustainability within their problem space and practical implementation, although it relates to sustainability’s social, environmental, and economic dimensions. This can help understand sustainability issues internally or for a more expansive understanding and application of methods for sustainability-related problems.

DL and neural network models are major methods for addressing sustainability-related problems among these categories. This is evident from the keyword analysis, where DL is shown among the top keywords. DL models are specifically for prediction, forecasting, optimization, and classification. The data for these methods are IoT data, BD, production and manufacturing data, text data, structured data, temporal data, image data, and geological and satellite data. The methods are beneficial for where data is in large amounts and well managed. The social dimension of DL applications includes challenges related to sustainable urbanization, education, health, and healthcare automation. The other major application is the environmental dimension for smart and sustainable agriculture, soil management, waste management, sustainable energy usage, wetland and ecosystem management, crop yield, and plant disease prediction. The economic dimension applies to resource optimization, smart and sustainable manufacturing, inventory control, financial crisis prediction, efficient scheduling, resource allocation, predictive maintenance, condition monitoring, and energy-efficient production. The DL and neural network models include ensemble models, convolutional neural networks, deep transfer learning models, Gaussian and Bayesian neural networks, recursive neural networks, recursive support vector neural networks, deep reinforcement learning models for classification, regression, forecasting, and unsupervised learning models with deep encoder-decoder architectures.

It is important to distinguish DL models from ML models because they are data- and resource-intensive; other than that, the lack of explainability of models cannot fit them into certain applications, particularly those related to critical health decision-making or policy formulation where transparency is crucial. TinyML is an emerging trend as edge computing provides ML-as-a-Service (MLaaS) to IoT devices. The resources are limited for edge computing, and using deep models cannot be useful for response time and accuracy. However, the other ML models can be useful for tinyML [361]—other than that, the other methods are more useful where the data is sparse and requires fewer resources.

The other supervised ML methods are widely used for various sustainability-related questions. It is shown that they cover a wide range of sustainability applications for pattern recognition, prediction, and classification. The most applied approaches for sustainability tasks are tree-based models, including decision trees, random forest, and gradient boosting. The other methods are Support Vector Machine (SVM), regression models (linear, polynomial, beta, Bayesian, and support vector regression), and ensemble models. Other than predicting various social, environmental, and economic questions and feature selection shown in Table 10, various studies utilize these methods to evaluate different sustainability tasks, socio-economic factors, and environmental challenges. For example, Molina-Gómez et al. [218] use different sustainability indicators to collect data to predict different levels of sustainable development of the urban ecosystem. This approach can be useful for identifying the quality of life in urban ecosystems and policy-making decisions for disadvantaged societies. The other application is sustainability performance assessment with SVM for large-scale decision-making for inclusive decisions with incomplete data [126]. Yigitcanlar and Cugurullo [357] discuss key aspects of the sustainable adoption of AI that require significant improvement for sustainable urbanism. Similar aspects are needed to design for sustainable adoption.

The unsupervised models are applied for data exploration, formulation of hypotheses, dimensionality reduction, latent variable modeling, anomaly detection, and clustering. The unsupervised methods are independently applied or in conjunction with other data-driven and machine-learning approaches. Some examples where it is integrated with other methods: Nilashi et al. [231] use clustering and supervised learning methods to predict the overall sustainability performance of different countries. Tayal et al. [317] discuss a staged approach, including data envelopment analysis, meta-heuristic approach, and k-means clustering to optimize sustainable facility layout design (shop floor, manufacturing unit), maximizing performance and minimizing operating costs.

The other methods are forecasting and time series-based models, which require estimating predictions based on historical temporal data and are useful for anticipating future events to reduce uncertainty, analyzing trends, optimizing resources, and taking proactive actions. The various applications are forecasting energy and other resources usage, load management, and forecasting values or patterns of various social, climate-related, and economic variables. These forecasting models can be applied for short and long-range forecasting. Clark et al. [59] discuss the need for iterative near-term ecological forecasting that maximizes ecological relevance to society,similar situations can be realized in various other social and economic cases, where the availability of BD can make iterative forecasting approaches possible for effective and adaptive decision making which can critically evaluate and update models based on new understanding. BD for forecasting is categorized into three categories: user-generated, device, and log data. Successful utilization of BD in forecasting demands a systematic approach: expertise in method development/selection (including statistical, ML, and hybrid methods), data processing, and domain knowledge [313].

ML model applications include transportation, food security, urbanization, agriculture, governance and public services, poverty, education, water, sustainable goal evaluation and SC applications, energy optimization, manufacturing, sustainable material properties, sustainable power supply, and green technology adoption. The designing of efficient AI algorithms is a challenge due to the open-ended complexity of problems related to sustainability. As more and more data are available in different domains, a wide range of applications of AI-based methods for sustainability problems become visible. From the implementation point of view, data quality, dimensionality, online effect, interpretability, accuracy, generalization, hyperparameter tuning, model updating, cost, and resources involved in developing models are key factors for appropriate ML model selection for ML methods. To ensure the robustness and sustainability of AI, it is imperative to examine how AI can adapt to external changes for long-term sustainability and identify the model development requirements that consider temporal effects, data characteristics, hyperparameter selection, continuous model updating, interpretation, and explainability.

One of the important cases is the multi-dimensional consideration of the problems in all three directions of sustainability. For instance, if an organization solely targets reducing expenditure and maximizing profit by minimizing energy usage and carbon footprint within its operations, it may inadvertently overlook other conflicting objectives across different dimensions of sustainability. This limited approach might yield immediate results by promoting the adoption of alternative energy sources. However, such alternatives could introduce challenges. Transitioning to these sources might necessitate utilizing resources with a substantial carbon footprint and additional maintenance requirements, potentially compromising long-term sustainability. Moreover, the cumulative effects could trigger broader ecological repercussions if multiple organizations embrace similar energy solutions. A widespread shift towards specific energy storage or generation methods could escalate the demand for particular materials, resulting in unforeseen environmental consequences. Therefore, the long-term effect must be understood by data-driven methods in maintaining sustainability.

Several multi-objective optimization methods can be applied to various social, environmental, and economic factors. The optimization methods are largely applied to complex problems to search for the best feasible solution, maximizing sustainability objectives considering problem constraints [272]. Metaheuristic, Pareto, scalarization, dominated, non-dominated, stochastic, and simulation methods can be applied to sustainability problems [82, 109, 136].

Multi-criteria decision-making is another area of methods applied as a decision support system for complex technological, social, and economic systems. Policymakers, managers, investors, and academicians can use these decision-making methods to identify and prioritize key indicators. Multi-criteria decision-making methods with fuzzy approach and simulations are applied to address various aspects of sustainability in multiple sectors. These include building and urban design, SCM, production, transportation, air pollution and waste management, and sustainable project selection. These methodologies assist in addressing complex decision-making strategies encompassing multiple conflicting criteria underneath uncertainty. This highlights the key role of these methods for optimization and decision-support tools for identifying sustainable solutions in diverse fields. The data sources for these studies include surveys, expert groups, observational studies, databases, or experimental design, depending on the specific context and requirements of the decision problem by identifying influential, dependent, and conflicting variables.

Sustainable big data and analytics

With the recent advancement in ICT and its widespread application, data can easily be gathered from various sources, including governmental, industrial, transactional, and urban data. “Datafication” has become an important factor in many sectors. It describes the current trend of defining phenomena, processes, and operations by turning them into digitized data to analyze and extract information [41]. The extracted information offers a huge potential to reinforce social, environmental, and economic sustainability. The prospect of leveraging BD resources to improve sustainability is a topic of research in many domains, including but not limited to sustainable agricultural SCs, sustainable health care, sustainable organization performance, innovation, sustainable urban planning, and smart cities.

Multiple studies discuss using BD approaches for sustainability applications and their potential to improve sustainability. Kamble et al. [150] highlight the potential of data analytics to augment the sustainability of the agri-food SC. For instance, the author discussed how descriptive analytics, namely the life cycle assessment tool, is widely used to address environmental concerns, analyze the misuse of resources leading to food waste, and design short SCs that increase the involvement of farmers and contribute to sustainable community development. Suvarna et al. [311] examine how “industrial data” can be leveraged by manufacturing companies for multiple purposes, including real-time process monitoring, data-driven process control and optimization, and decentralized manufacturing. This falls in line with the improvement of sustainability. From an economic perspective, improving the efficiency of manufacturing processes through control and optimization improves the company’s financial profitability and ensures its economic sustainability. The improved process efficiency also significantly impacts energy consumption and resulting emissions, strongly contributing to environmental sustainability. Moreover, by allowing decentralized manufacturing, the exploitation of industrial data can contribute to a more even geographical distribution of job opportunities and hence empower social sustainability.

Similarly, BD plays a key role in health care. The literature suggests several BD tools that could be implemented in a digital healthcare system and allow for the exploration of hidden patterns, early detection of diseases, and reduction of the required time and cost of analysis [360]. “Open Data” platforms are another tool that strongly contributes to sustainability goals. “Open data” platforms, where data can be shared between multiple parties, can lead to highly granular information content and allow all involved parties to benefit from it. For instance, the development of open-data healthcare platforms where citizens can share their medical data can contribute to better knowledge, diagnosis, and treatment of diseases. Moreover, locally or globally, sharing information from different population sections can help researchers better understand inequities, highlight them, and address them with dedicated healthcare interventions [174]. However, even though BD can offer lucrative opportunities in many fields and help reinforce sustainability, implementation comes with critical challenges. The challenges range from data collection to storage, processing, governance, privacy, and accessibility and pose questions about how sustainable BD approaches are from all social, environmental, and economic aspects [337].

H. Zhang et al. [364] highlight the difficulties service design teams face in fully understanding and exploiting BD due to a lack of big data analytic capabilities (BDAC). In line with that, excessive investment in BD resources alone without parallel investment in BDAC can hinder sales growth and negatively affect companies’ sustainable growth [113]. Apart from the negative implications of the financial sustainability of companies, the perspective of the knowledge gap hypothesis, which posits that significant investments in BD resources and analytics are required, can also be detrimental to social sustainability. In line with the knowledge gap hypothesis, when there is a rise in the introduction of mass media information into social systems, communities or countries with a higher socio-economic status typically acquire this information more quickly than their counterparts with lower socio-economic status. This results in an amplification of the gap between them rather than a reduction [234]. Applying this hypothesis to BD, economically advanced regions will likely have the capacity to swiftly make substantial investments in harnessing BD and developing essential skills, outpacing the progress of underdeveloped regions. Such a scenario could reinforce a monopoly on BD technologies, ultimately affecting social or economic aspects of sustainability.

In addition, the potential offered by open data platforms is opposed by serious privacy concerns, which in turn raise questions about the lack of “Data Governance” standards to regulate the way data is collected and processed, as well as define who is accountable for the decision making resulting from data usage [258]. This lack of governance can lead to data usage that contradicts social and environmental sustainability goals, namely climate protection and reducing inequalities. Within the social context, some researchers discussed the thriving data broker industry in which data is used for purposes it was never intended for, including “to predictively profile, socially sort, behaviorally nudge, and regulate, control and govern individuals and the various systems and infrastructures with which they interact” [166]. This trend is alarming in terms of its implications for individual privacy. Regarding the environmental aspect, the amount of data generated from various sources requires energy-intensive computational resources to collect, store, and conduct analytics. This could potentially negate the positive effect of BD and its analytics on sustainability. These powerful computational resources require mining rare materials for their manufacturing, consuming high energy levels for processing and cooling, and generating waste when discarded, raising questions about their sustainability.

Moreover, the ease and low cost of data collection and its potential financial benefits have encouraged organizations to focus on accumulating vast amounts of data in their databases to commercialize it. In the absence of governance to regulate how data is collected, this can lead to “data obesity,” in which case redundant data might be collected from multiple sources, thus increasing the size of datasets without contributing to their information content [213]. To counteract the effect of data redundancy, allowing for more energy-sustainable data sharing and computing, the scientific community is reverting to “data aggregation and fusion” techniques, which aim to reduce data size without affecting its information. Data aggregation and fusion encompass a variety of methods, including redundant data elimination, data compression, in-network processing, and data sampling and prediction [52]. The challenge remains, however, in minimizing the data quality loss when reducing its size, which is a prominent area of research. For instance, Pielli et al. [253] propose an RL technique to choose the compression rate that maximizes data quality and adheres to the energy constraint.

Even though research seeks to optimize techniques that reduce data size without affecting its quality, a correlation between data size and quality still exists in certain cases. This conflict arises between the financial rewards of data approaches and their environmental impact. Commercial institutions face the challenge of deciding between financial gain and environmental sustainability due to the absence of proper data governance or a framework defining responsibilities. The literature shows that exploiting BD to enhance sustainability and analytics still faces significant challenges. While the potentials presented are promising, the different social, environmental, and economic aspects are interconnected and, at times, can be inversely proportional. Therefore, implementing such approaches can have differing impacts on each aspect of sustainability. The field of BD is an active research topic with ongoing efforts by scholars to refine the analytic methods and develop frameworks that can mitigate the potential negative implications of BD approaches.

Sustainable AI characteristics and challenges

The IoT and AI have the potential to support prevailing sustainability measures like the circular or sharing economy. Fraga-Lamas et al. [95] explicitly mentioned that IoT could pave the way for more sustainability in various sectors in operations, maintenance, and processes and areas addressed in the UN’s SDGs, like water distribution. Bachmann et al. [30] delivered a comprehensive work about the contribution of data-driven technologies (including IoT) to achieving the SDGs. Sætra [274] emphasizes the different levels AI might influence in the context of the SDGs. As an example, it is pointed out that positive effects for a country (meso level) might create negative effects on other countries (macro level) and raise tensions within the country (micro level). Therefore, Sætra [274] describes AI not as a single technology but as something connected to various other technologies that may simultaneously trigger positive and negative effects.

On the one hand, IoT and AI can serve as facilitators for sustainability, but paradoxically, the IoT sector itself has a high carbon footprint [95]. Also, Schwartz et al. [285] claim that ML-based text and image generators have improved in quality over the last few years. However, this AI research has become increasingly “Red AI,” meaning it seeks to improve accuracy by using massive computational power, which is costly, environmentally unfriendly, and exclusive. Furthermore, Schwartz et al. [285] propose that a “Green AI” should treat efficiency as a primary evaluation criterion alongside accuracy, and they suggest reporting the number of Floating-Point Operations (FPO) required to generate a result as an indicator of how green an AI-based result is. An approach to optimize the energy consumption of Deep Neural Networks (DNN) is presented by Liu et al. [198,199,200] through a combination of deep and spiking neural networks, which lead to a six times lower energy consumption. Another approach, the optimization framework called “Zeus,” is presented by You et al. [359]. This framework proposes to decouple the optimization of batch size and Graphics Processing Unit (GPU) power limit. It uses an online exploration–exploitation approach based on multi-armed bandit and just-in-time energy profiling to navigate the trade-off between energy consumption and performance optimization. The work by S. Choi et al. [54] focuses on energy saving when using multi-GPUs for DNN training with their proposed framework “EnvPipe.” This framework should save energy by using multiple GPUs with pipeline parallelism while providing high accuracy and performance. However, this highlights that the challenges are multifaceted, encompassing identifying appropriate application fields for AI and IoT to enhance a business’s sustainability and determining the optimal level of AI and IoT utilization.

Besides the intensity of IoT use, Dunn et al. [80] are especially addressing the vulnerability of IoT systems. They claim that ML technology, often the basis for IoT systems, could suffer from data poisoning or cyber-attacks, which create false assumptions, results, or outcomes and may harm the application field. In smart cities, for example, AI can support efficiencies in decision-making, infrastructure assessment, post-disaster reconnaissance, connected urban mobility, or service agent chatbots, to name a few [358]. However, its use is reluctant due to the technocratic use of AI and its not being implemented sustainably. The main focus has been on increasing efficiency, but a sustainable implementation also needs a social and environmental equitable point of view [358]. The challenge of sustainability in the product service system is addressed by Xinyu Li et al. [193]. Their article emphasizes integrating sustainability considerations into Product-Service Systems (PSS) design and development, particularly in a cyber-physical environment. It underscores the lack of existing studies that see cyber-physical resources as a whole regarding sustainability rather than just physical materials and components. As cyber-physical systems often use AI or IoT technology, Xinyu Li et al. [193] propose a data-driven reversible framework that extends the traditional scope of resource management to create Sustainable Smart PSS. This framework incorporates a four-step context-aware process that involves requirement elicitation, solution recommendation, solution evaluation, and knowledge evolution to support decision-making and optimization throughout the product lifecycle.

Apart from the technological point of view, a challenge of AI is presented by How et al. [122]. The authors emphasize the critical role that AI can play in understanding and promoting sustainable development. It specifically focuses on how AI can be made more accessible and user-friendly to people who do not have a computer science background. How et al. [122] introduce a novel, human-centric probabilistic reasoning approach that democratizes AI by allowing non-computer scientists to use AI to analyze socio-environmental data. Wilson and van der Velden [349] focus on the public sector and claim that ethics, explainability, responsibility, and accountability are important aspects of analyzing the societal impacts of AI. These concepts alone do not support regulating and implementing AI in the public sector. Wilson and van der Velden [349] explored the concept of “sustainable AI” in their work to address this gap by aligning the research on sustainable development with that on public sector AI. A conceptual model identifying five boundary conditions, diversity, learning capacity, self-organization, common meaning, and trust, can assist in public sector decision-making about AI governance [349]. There are still challenges when using AI in smart cities, production, or the public sector. Scholars work to make AI more sustainable,they review current approaches, derive new possible pathways, and propose frameworks, models, and concepts to make AI more sustainable.

Sustainable human AI ecosystems

The different data-driven and AI methods for sustainable applications are applied at different levels of decision-making where the outcome of these methods is involved for subsequent levels considering several factors, including social, environmental, and economic aspects and the local, regional, and global constraints. Different methods can be useful for problems at different levels, from large data-intensive AI models to tiny or small data-related decision-making models, ML, and analytics. It is, therefore, useful to know the effectiveness and power of different methods at different levels of decision-making given the various method-related, data-related, and other application-related constraints, explainability, and impact on decisions.

A systematic framework and a hybrid approach to decision-making are critical and show better performance, context awareness, and improved accuracy [282]. The hybrid approach involves humans and different AI and data-driven methods, and its aim should be to maximize the impact of decision-making towards sustainability goals. The framework must ensure who will make the final decisions, and the allocation of responsibility for decision-making is an important consideration. While AI can significantly assist decision-making, there should always be a mechanism to ensure that human judgment and values can supersede AI-generated decisions when necessary.

The other challenge is the implications that can originate from large-scale interactions between “AI” and “humans.” The future of AI will be about AI not working in isolation but evolving as multi-agent systems, and AI (or multi-agent)-human ecosystem where humans and AI systems continually interact and learn from each other lead to an extensive shift in decision-making processes. This effect would generate a social impact and affect the decision-making compass of companies, policy-making bodies, and other decision-making entities, subsequently influencing sustainability-related aspects. Key considerations include the consequences of bias in these decisions [310] or error propagation initiated either by humans or AI [72]. These errors have the potential to propagate through the system and compound due to repeated interactions. The other interesting part is that AI and human interactions as a Social and Technological System (STS) lead to unknown emergent behavior from the aggregated outcomes that can lead to conformism and potential unanticipated and undesirable consequences [246]. Pedreschi et al. [246] emphasize the design of next-generation AIs with the “complexity-informed perspective” abided by the sustainability goals where individual and collective concerns are addressed effectively.

Conclusion and future research

The study has conducted four main analyses, which include an overview of publications, an analysis of research articles across multiple countries that showcase collaborative research actions in the context of sustainability tasks and data-driven methods, keywords and co-occurrence analyses, and topic modeling. Collaborative research highlights the substantial lack of collaboration in underdeveloped regions and is to be emphasized further for global sustainability goals. The frequency of emerging concepts, application areas, and key technologies has been highlighted by analyzing keywords, particularly concerning AI and data-driven approaches that are gaining prominence. The modules with keywords in the co-occurrence network reveal common interconnected concepts and their social, environmental, and economic dimensions, highlighting the relevance of the complexity of different data-driven applications. The topic analysis has also highlighted the major research area related to sustainability and data-driven applications. These findings from the literature have further driven us to explore the different AI and data-driven methods applied to different types of applications. We have found that nine categories of methods are applied for different cases and applications for different levels of decision-making and their importance. It is important to understand that a single method cannot show its importance in decision-making. However, different methods have their utility and shortcomings for various sustainability-related problems, covering different aspects of sustainability and the scope of applications in engineering, basic research, health and medicine, business and policy-making, and governance. Therefore, their scope and power in multi-level and multi-dimensional decision-making must be understood before using different methods for decision-making, given the data and methods-related constraints and their ability to provide meaningful and reasonable explanations.

It is important to remember that different decision-making processes are not single, isolated decisions but multi-level and multi-dimensional processes requiring different methodologies. Therefore, it is crucial to systematically integrate these methods to maximize their ability to provide accurate and long-term insights into sustainability. Additionally, BD and AI will play a significant role in decision-making. With the expansion of IoT and other technologies, BD and AI will be widespread, posing several sustainability-related challenges. AI continues to evolve as a multi-agent system and a human-AI ecosystem where new challenges will arise. We must be aware of them to apply AI to address sustainability challenges effectively and adopt sustainable AI characteristics in both development and decision-making processes. Additionally, it is imperative to remain informed about the personal and collective responsibilities associated with its use, its future implications, and potential challenges.

Limitations and future research

Our analysis may have limitations, such as relying on a single database for selecting research papers, which may overlook certain published papers and lead to incomplete results and understanding of some topics. Additionally, selecting keywords that define social, environmental, and economic dimensions may lack in-depth understanding as the research area is multidisciplinary and can only be based on limited keywords. Additionally, including a more detailed list of AI and technology-related keywords in the search criteria can enhance the results’ quality and relevance. Furthermore, topic analysis based on only the abstract, title text, and keywords may limit understanding of the subject matter. In contrast, a more comprehensive approach would involve an in-depth analysis of the entire text. Fine-tuning approaches, including different embeddings, dimension reduction methods, and clustering methods, and aggregating different types of analysis could further improve results. Employing multiple semantic analysis methods and incorporating domain understanding for topic selection, interpretation, and identifying key representative documents is advisable to enhance result stability.

Future research could focus on data-driven methods for sustainable BMI, sustainable SC analysis, and developing a framework that systematically integrates different methods for different levels of decision-making in BMI for robust outcomes and complementing all aspects of sustainability.

Appendix

See Tables 11, 1213 and Figs. 9, 10.

Table 11 Representative keywords from c-TF-IDF matrix for different topic segments from BERTopic modelling

Table 12 Different ML and data-driven methods applied to different sustainability problems

Table 13 Keywords in different modules in co-occurrence network

Fig. 9

UMAP-based 2-dimensional projection of document embeddings and BERTopic-identified topics (color-coded)

Fig. 10

N-Statistic values comparing differences between topic distributions based on the 2D UMAP projection of document embeddings and the clusters (topics) obtained from 'hdbscan' in BERTopic